Anthropic dropped Claude Haiku 5.5 on October 7, 2026, with a pricing structure that looks too good to be true until you hit the token limit. The model, identified as claude-haiku-5-5, costs just $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100,000 tokens. That is a 90% price cut compared to Haiku 4.5. However, cross that 100k threshold and the rates jump to $0.50 input and $2.50 output, reducing the savings to only 50%. This binary pricing tier turns every request into a potential budget trap if your application doesn't track prompt size rigorously.

The Hidden Costs of Effort and Tokenization

Beyond the cliff, two other factors distort the 'cheap' narrative. First, the new tokenizer consumes approximately 30% more tokens for the same text compared to Haiku 4.5. A cheaper per-token rate does not automatically mean a cheaper total bill if you are processing 30% more units. Second, Haiku 5.5 introduces an effort dial with five settings: low, medium, high, xhigh, and max. The API defaults to medium, but higher effort levels increase the computational cost without published multipliers. Developers cannot simply assume the list price reflects the final invoice when using high or max effort settings.

Building a TypeScript Price Gate

To navigate this complexity, developer Bobby Hall Jr. released a TypeScript utility that acts as a pre-flight price gate. The code does not call the API; instead, it accepts token counts and task types to return an ALLOW, REVIEW, or REFUSE decision. It enforces strict rules: only classify, extract, route, summarize, and compact tasks are allowed at low or medium effort. Agentic coding and computer-use tasks trigger a REVIEW status, even on short prompts, because Haiku 5.5's Terminal-Bench score of 39.2% lags significantly behind Sonnet 5.5's 70.6%. The gate also flags any prompt exceeding 100,000 tokens as REVIEW, forcing developers to acknowledge the 5x price increase before execution.

Key Takeaways

  • The 90% savings claim only applies to prompts under 100,000 tokens; longer contexts see only 50% savings.
  • The new tokenizer uses ~30% more tokens than Haiku 4.5, partially offsetting the lower per-token cost.
  • The effort dial (low to max) changes billing without published multipliers, requiring careful evaluation for high-effort tasks.
  • Sonnet 5.5 cache read prices were also cut to $0.10 per million tokens on October 7, 2026.

The Bottom Line

Haiku 5.5 is a powerful tool for simple tasks but a financial landmine for complex workflows; you must build your own price gate before sending requests.